Module 06 — Economics and model choice

Time: 3 hours · Format: 60 min theory + 120 min practice

Theory

1. What goes into the cost

The price is quoted per 1 million tokens and splits into:

Type	What it is	Usually
input	your context	expensive
output	the model's answer	3–5× more expensive than input
cache write	writing to the cache	≈ input
cache read	reading from the cache	5–10× cheaper than input

Session formula:

cost ≈ (input_tokens × price_in + output_tokens × price_out) / 1_000_000

But in an agent loop input is multiplied by the number of steps: every step resends the entire context. 10 steps with a 50k context means 500k input tokens.

2. Cache — the main lever for savings

Providers cache the stable prefix of a request. If the structure is always:

[instructions] [project files] [question]

then on repeated steps the prefix is read from the cache, and you pay pennies.

What kills the cache:

changing instructions mid-session;
rearranging files;
inserting the current time/random data at the start;
starting every step "from a clean slate".

3. Latency

It matters for UX, but not only:

a reasoning model thinks for 30–120 seconds before the first token;
on short edits that is annoying;
on architectural tasks it pays off.

4. Model classes

Class	Trait	What for
fast / mini / flash	cheap, fast	renames, tests by example, search
balanced	mid-range	everyday development, refactoring
frontier + reasoning	expensive, thinks	architecture, hard debugging, migrations
long-context	500k–2M window	analyzing large repos/logs

Rule: 80% of tasks are solved by a cheap model if you have tests and a good spec.

5. How to choose in practice

1. Is there a check?     no  → oracle first, then the model
2. Routine task?         yes → fast
3. Needs thinking?       yes → reasoning
4. Context > 200k?       yes → long-context is mandatory
5. Many tool-calls?      yes → tool-calling reliability matters, not "smarts"

Another nuance: the model's temperament. One likes writing tests, another dives straight into the code, a third is too verbose. You need 2–3 favorite models per task type, not one "best" one.

6. Budget and limits

Set provider limits (a monthly budget).
Cap max_tokens for routine tasks.
Keep an eye on opencode stats.
A single "vibe session" at 2M tokens can cost as much as a day of work — that's normal if the result is worth it, and a disaster if it was a headline edit.

7. When an expensive model is cheaper

A paradox: a frontier model over 3 iterations can cost less than a mini model over 15 iterations plus your time on review. You have to count total cost: tokens + your time + bugs.

Comprehension checklist

Why is input multiplied by the number of steps in an agent loop?
What kills the cache and how do you fix it?
Name three tasks where a cheap model is objectively better than an expensive one.
How do you compute the "total cost" of a task?

Practice

See practice.md.